idkiro / attack-imagenet Goto Github PK

where is the implement...

When I run your code, I get the following error:
TypeError: einsum() takes 2 positional arguments but 3 were given

def forward(self, x):
    if self.embed:
        theta = self.conv_theta(x)
        phi = self.conv_phi(x)
    else:
        theta = x
        phi = x
    n_in, H, W = list(x.size())[1:]
    if n_in > H*W or self.softmax:
        f = torch.einsum('niab,nicd->nabcd', theta,phi)
        if self.softmax:
            shape = f.size()
            f = f.view(-1, shape[2]*shape[3], shape[2]*shape[3])
            f = f / math.sqrt(self.channel/2)
            f = nn.functional.softmax(f, dim=-1)
            f = f.view(shape)
        f = torch.einsum('nabcd,nicd->niab', f, x)
    else:
        f = torch.einsum('nihw,njhw->nij', phi, x)
        f = torch.einsum('nij,nihw->njhw', f, theta)
    if not self.softmax:
        f = f / (H*W)
        
    y = self.bn(self.conv(f))
    return x + y

Please help me solve this puzzle, thank you!

When I run simple_attack.py.
There is an error:

simple_attack.py 85
model.cuda()

module.py 260 cuda
return self._apply(lambda t: t.cuda(device))

module.py 187 _apply
module._apply(fn)

module.py 187 _apply
module._apply(fn)

module.py 187 _apply
module._apply(fn)

module.py 193 _apply
param.data = fn(param.data)

module.py 260
return self._apply(lambda t: t.cuda(device))

init.py 161 _lazy_init
_check_driver()

init.py 75 _check_driver
raise AssertionError("Torch not compiled with CUDA enabled")

AssertionError:
Torch not compiled with CUDA enabled

How to fix it?

请问您的代码生成的对抗样本的可迁移性是如何评估的呢？所攻击的黑盒模型是什么呢？

Tensorflow的模型如何转为PyTorc权重并使用的，代码可否发布一下。

When I run your code, I get this error:

RuntimeError: cuDNN error: CUDNN_STATUS_NOT_SUPPORTED. This error may appear if you passed in a non-contiguous input.

The error occurs when I run line 83 "loss.backward()" in attacker.py, I don't know what's going on here.
Maybe you know how to solve it?

is_better = loss < best_loss; best_loss[is_better] = loss[is_better]; best_delta[is_better] = delta.data[is_better]
Would you please tell me the meaning of this code?The following error was reported while running:
“too many indices for tensor of dimension 0”

请问为什么要给标签添加偏移量啊，是原始数据已经做过处理了吗，这个偏移量使我产生了类型的错误subscriptable，如果不予理会，会事得分降低吗。